XFS does its own data writeout from several places using xfs_flush_pages().
When this writeout is racing with flusher thread writing the same inode the
performance gets bad because flusher thread submits writes as normal WRITE
commands while xfs_flush_pages() submits them as WRITE_SYNC (as it uses
filemap_fdatawrite_range()) and CFQ does not merge such requests. So we end up
with relatively small requests from the flusher thread and sync interleaved.
Short excerpt from blktrace:
254,16 2 16949 103.786278461 25233 Q W 54479888 + 424 [flush-254:16]
254,16 4 20662 103.786386751 25232 Q WS 54492920 + 1024 [dd]
254,16 4 20665 103.786444241 25232 Q WS 54493944 + 104 [dd]
254,16 2 16952 103.786533451 25233 Q W 54494048 + 288 [flush-254:16]
254,16 4 20668 103.787295241 25232 Q WS 54494336 + 1024 [dd]
254,16 4 20671 103.787332801 25232 Q WS 54495360 + 216 [dd]
254,16 2 16955 103.789062500 25233 Q W 54495576 + 1024 [flush-254:16]
(actually, I was observing even smaller requests on a different HW which isn't
available to me now).
Although we cannot fix all the cases (e.g. when fsync is called), we can use
filemap_flush_range() for the case when xfs_flush_pages() is called with
XFS_B_ASYNC flag which fixes the problem at least for some common cases like
the first round of sync writeback or flushing of data during truncate or after
file is closed.
Signed-off-by: Jan Kara <jack@xxxxxxx>
---
fs/xfs/linux-2.6/xfs_fs_subr.c | 6 ++++--
1 files changed, 4 insertions(+), 2 deletions(-)
diff --git a/fs/xfs/linux-2.6/xfs_fs_subr.c b/fs/xfs/linux-2.6/xfs_fs_subr.c
index ed88ed1..6f33164 100644
--- a/fs/xfs/linux-2.6/xfs_fs_subr.c
+++ b/fs/xfs/linux-2.6/xfs_fs_subr.c
@@ -70,10 +70,12 @@ xfs_flush_pages(
int ret2;
xfs_iflags_clear(ip, XFS_ITRUNCATED);
+ if (flags & XBF_ASYNC) {
+ return -filemap_flush_range(mapping, first,
+ last == -1 ? LLONG_MAX : last);
+ }
ret = -filemap_fdatawrite_range(mapping, first,
last == -1 ? LLONG_MAX : last);
- if (flags & XBF_ASYNC)
- return ret;
ret2 = xfs_wait_on_pages(ip, first, last);
if (!ret)
ret = ret2;
--
1.7.1
--
To unsubscribe from this list: send the line "unsubscribe linux-fsdevel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
